Shoeprint Alignment for Feature Extraction

Image Database

• We are working with the EverOS 2D scanner images
• This collection has scans from around 150 pairs of shoes
• Each shoe was scanned four times, with four replicates each time
• There are four different sizes of two different brand/models

Main Objective

To develop an automatic algorithm to match two scans corresponding to the same shoe based on analysis of cross-section

In order to do this we need:

1. Align images so we can identify the same cross-sections on all the shoes
2. Identify the areas with the most potential for discrimination
3. Extract cross-section signatures in such areas
4. Study the correlationn of cross-section on both match and non-match pairs
5. Use the cross-section signatures to construct informative features
6. Fit and asses statistical models

Shoeprint image registration/alignment

• After a variety of attemps, the best way to proceed is using the MATLAB function “imregister”
• Even with the orignal size of the images (approx. 4000 x 2000 pixels) an alignment takes up to one minute on a regular laptop
• The MATLAB method is robust enought to align the images even when the positions are quite different
• Just few times (less than 10% of the cases) it’s neceesary to re-run the function with an adjusted parameter

Alignment Example

Next Steps

• It’s necessary to align all the images (different sizes and brand/models) to a similar position
• We need this to be able to talk about the same cross-section regarless of size and brand/model
• Then we will be able to determine the areas of the shoesole with the most potential for discrimination

Example of a Cross-Section